Research Tool Captures Fine-Grained Changes in Jupyter Notebooks to Advance Data Science Development Practices

Computational notebooks like Jupyter have become essential infrastructure for data science and machine learning work, yet they remain largely understudied compared to traditional software projects. A new research paper addresses that gap by introducing tools and data to illuminate how developers actually work within these environments.

Capturing Real Development Workflows#

The researchers built instrumentation to track cell-level changes in Jupyter notebooks as they occur during active development. They collected data from 20 developers with varying levels of expertise working on data analysis and machine learning tasks, gathering more than 100 hours of session data. This resulted in a dataset spanning 2,655 cells and 9,207 individual cell executions—creating a rich record of how notebooks evolve in practice.

Key Findings on Iteration Patterns#

Analysis of the collected logs revealed that code iteration modifications constitute a substantial portion of changes between cell executions. By classifying the types of edits developers make, the team identified concrete patterns in how exploratory and experimental workflows unfold. These granular observations move beyond high-level snapshots to show the moment-to-moment decisions developers make while debugging, refining, and validating their work.

Implications for Future Research#

The dataset opens new research directions for software engineering studies in computational notebooks. Prior work on code change analysis has yielded practical advances in refactoring, security, and code completion tools for traditional development—areas where detailed logs of developer behavior proved invaluable. The authors argue that similar fine-grained study of notebook development can drive similar innovations tailored to how data scientists actually work.